57
An Editorial Network for Enhanced Document Summarization
Edward Moroshko∗
Electrical Engineering Dept. Technion – Israel Institute of Technology
Haifa, Israel
edward.moroshko@gmail.com
Guy Feigenblat, Haggai Roitman, David Konopnicki
IBM Research AI Haifa University Campus
Haifa, Israel
{guyf,haggai,davidko}@il.ibm.com
Abstract
We suggest a new idea ofEditorial Network–
a mixed extractive-abstractive summarization
approach, which is applied as a
post-processing step over a given sequence of extracted sentences. We further suggest an effective way for training the “editor" based on a novel soft-labeling approach. Using the CNN/DailyMail dataset we demonstrate the effectiveness of our approach compared to state-of-the-art extractive-only or abstractive-only baselines.
1 Introduction
Automatic text summarizers condense a given piece of text into a shorter version (the summary). This is done while trying to preserve the main essence of the original text and keeping the generated summary as readable as possible.
Existing summarization methods can be classified into two main types, eitherextractiveor abstractive. Extractive methods select and order text fragments (e.g., sentences) from the original text source. Such methods are relatively simpler to develop and keep the extracted fragments untouched, allowing to preserve important parts, e.g., keyphrases, facts, opinions, etc. Yet, extractive summaries tend to be less fluent, coherent and readable and may include superfluous text.
Abstractive methods apply natural language paraphrasing and/or compression on a given text. A common approach is based on the encoder-decoder (seq-to-seq) paradigm (Sutskever et al., 2014), with the original text sequence being encoded while the summary is the decoded sequence.
∗
Work was done during a summer internship in IBM Research AI
While such methods usually generate summaries with better readability, their quality declines over longer textual inputs, which may lead to a higher redundancy (Paulus et al.,2017). Moreover, such methods are sensitive to vocabulary size, making them more difficult to train and generalize (See et al.,2017).
A common approach for handling long text sequences in abstractive settings is through attention mechanisms, which aim to imitate the attentive reading behaviour of humans (Chopra et al.,2016). Two main types of attention methods may be utilized, eithersoftorhard. Soft attention methods first locate salient text regions within the input text and then bias the abstraction process to prefer such regions during decoding (Cohan et al.,
2018; Gehrmann et al., 2018; Hsu et al., 2018;
Nallapati et al., 2016; Li et al., 2018;Pasunuru and Bansal,2018;Tan et al.,2017). On the other hand, hard attention methods perform abstraction only on text regions that were initially selected by some extraction process (Chen and Bansal,2018;
Nallapati et al.,2017;Liu et al.,2018).
Compared to previous works, whose final summary is either entirely extracted or generated using an abstractive process, in this work, we suggest a new idea of “Editorial Network" (EditNet) – a mixed extractive-abstractive summarization approach. A summary generated byEditNetmay include sentences that were either extracted, abstracted or of both types. Moreover, per considered sentence,EditNetmay decide not to take either of these decisions and completely reject the sentence.
Summary extracted sentences
Extractor
Original document Abstracted sentence
Sentence Representation
Extracted sentence representation Document
representation
Abstracted sentence representation Fully connected
Fully connected
E A R
Selection representation
Sentence Edit Decision
d
Summary representation
Document sentences
ℎ →
Generated Summary
Abstractor
D
S’
[image:2.595.73.297.61.264.2]S
Figure 1: Editorial Network (EditNet)
by both state-of-the-art abstractive-only and extractive-only baselines.
2 Editorial Network
Figure 1 depicts the architecture of EditNet. EditNet is applied as a post-processing step over a given input summary whose sentences were initially selected by some extractor. The key idea behind EditNet is to create an automatic editing process to enhance summary quality.
Let S denote a summary which was extracted from a given text (document) D. The editorial process is implemented by iterating over sentences in S according to the selection order of the extractor. For each sentence inS, the “editor" may make three possible decisions. The first decision is to keep the extracted sentence untouched (represented by labelEin Figure1). The second alternative is to rephrase the sentence (represented by label A in Figure 1). Such a decision, for example, may represent the editor’s wish to simplify or compress the original source sentence. The last possible decision is to completely reject the sentence (represented by labelRin Figure1). For example, the editor may wish to ignore a superfluous or duplicate information expressed in the current sentence. An example mixed summary generated by our approach is depicted in Figure2
in the appendix, further emphasizing the various editor’s decisions.
Editor’s automatic summary:
E: what was supposed to be a fantasy sports car ride at walt disney world speedway turned deadly when a lamborghini crashed into a guardrail.A:the crash took place sunday at the exotic driving experiencea.
A:the lamborghini ’s passenger , gary terry , died at the sceneb.
R: petty holdings , which operates the exotic driving experience at walt disney world speedway , released a statement sunday night about the crash.
aOriginal extracted sentence: “the crash took place sunday at
the exotic driving experience , which bills itself as a chance to drive your dream car on a racetrack".
bOriginal extracted sentence: “the lamborghini ’s passenger ,
36-year-old gary terry of davenport , florida , died at the scene , florida highway patrol said"
Ground truth summary:
the crash occurred at the exotic driving experience at walt disney world speedway. officials say the driver , 24-year-old tavon watson , lost control of a lamborghini. passenger gary terry , 36 , died at the scene.
Figure 2: An example mixed summary (annotated with the editor’s decisions) taken from the CNN/DM dataset
2.1 Implementing the editor’s decisions
For a given sentences ∈ D, we now denote by seandsaits original (extracted) and paraphrased (abstracted) versions. To obtain sa we use an abstractor, whose details will be shortly explained (see Section2.2). Letes∈Rnandas∈Rnfurther
denote the corresponding sentence representations of se and sa, respectively. Such representations allow to compare both sentence versions on the same grounds.
Recall that, for each sentence si ∈ S (in order) the editor makes one of the three possible decisions: extract, abstract or rejectsi. Therefore, the editor may modify summarySby paraphrasing or rejecting some of its sentences, resulting in a mixed extractive-abstractive summaryS0.
Letlbe the number of sentences inS. In each stepi∈ {1,2, . . . , l}, in order to make an educated decision, the editor considers both sentence representations esi andasi as its input, together
with two additional auxiliary representations. The first auxiliary representation is that of the whole documentD itself, hereinafter denoted d ∈ Rn. Such a representation provides a global context for decision making. Assuming documentDhas
N sentences, let¯e= N1 N P s∈D
es. Following (Chen
and Bansal, 2018;Wu and Hu,2018a),dis then calculated as follows: d = tanh(Wd¯e+bd) , where Wd ∈ Rn×n and bd ∈ Rn are learnable
parameters.
the summary that was generated by the editor so far, denoted at stepiasgi−1 ∈Rn, withg0 =~0.
Such a representation provides alocal contextfor decision making. Given the four representations as an input, the editor’s decision for sentencesi ∈S is implemented using two fully-connected layers, as follows:
sof tmax(V tanh(Wc[esi, asi, gi−1, d] +bc) +b),
(1) where[·]denotes the vectors concatenation,V ∈
R3×m, Wc ∈ Rm×4n, bc ∈ Rm and b ∈ R3 are
learnable parameters.
In each stepi, therefore, the editor chooses the actionπi ∈ {E,A,R}with the highest likelihood (according to Eq.1), further denotedp(πi). Upon decision, in case it is either E or A, the editor appends the corresponding sentence version (i.e., either sei orsai) to S0; otherwise, the decision is
Rand sentencesiis discarded. Depending on its decision, the current summary representation is further updated as follows:
gi=gi−1+tanh(Wghi), (2)
where Wg ∈ Rn×n are learnable parameters,
gi−1 is the summary representation from the
previous decision step; and hi ∈ {esi, asi, ~0},
depending on which decision is made.
Such a network architecture allows to capture various complex interactions between the different inputs. For example, the network may learn that given the global context, one of the sentence versions may allow to produce a summary with a better coverage. As another example, based on the interaction between both sentence versions with either of the local or global contexts (and possibly among the last two), the network may learn that both sentence versions may only add superfluous or redundant information to the summary, and therefore, decide to reject both.
2.2 Extractor and Abstractor
As a proof of concept, in this work, we utilize the extractor and abstractor that were previously used in (Chen and Bansal, 2018), with a slight modification to the latter, motivated by its specific usage within our approach. We now only highlight important aspects of these two sub-components and kindly refer the reader to (Chen and Bansal,2018) for the full implementation details.
The extractor of (Chen and Bansal, 2018) consists of two main sub-components. The first
is anencoderwhich encodes each sentences∈D intoesusing an hierarchical representation1. The second is a sentence selector using a Pointer-Network(Vinyals et al., 2015). For the latter, let P(s)be the selection likelihood of sentences.
The abstractor of (Chen and Bansal, 2018) is basically a standard encoder-aligner-decoder with a copy mechanism (See et al., 2017). Yet, instead of applying it directly only on a single given extracted sentence se
i ∈ S, we apply it on a “chunk" of three consecutive sentences2 (se−, sei, se+), wherese−andse+denote the sentence
that precedes and succeedsse
i inD, respectively. This in turn, allows to generate an abstractive version of sei (i.e., sai) that benefits from a wider local context. Inspired by previous soft-attention methods, we further utilize the extractor’s sentence selection likelihoodsP(·)for enhancing the abstractor’s attention mechanism, as follows. LetC(wj)denote the abstractor’s original attention value of a given wordwjoccurring in(se−, sei, se+);
we then recalculate this value to be C0(wj) = C(wj)·P(s)
Z , with wj ∈ s and s ∈ {s e
−, sei, se+};
Z =P s0∈{se
−,sei,s e
+} P
wj∈s0C(wj)·P(s
0)denotes
the normalization term.
2.3 Sentence representation
Recall that, in order to comparesei withsai, we need to represent both sentence versions on as similar grounds as possible. To achieve that, we first replace sei withsai within the original document D. By doing so, we basically treat sentencesai as if it was an ordinary sentence within D, where the rest of the document remains untouched. We then obtainsai’s representation by encoding it using the extractor’s encoder in a similar way in which sentencesei was originally supposed to be encoded. This results in a representationasi that provides
a comparable alternative to esi, whose encoding
is expected to be effected by similar contextual grounds.
2.4 Network training
We conclude this section with the description of how we train the editor using a novel soft labeling approach. Given textS(withlextracted sentences), letπ = (π1, . . . , πl) denote its editing decisions
1
Such a representation is basically a combination of a temporal convolutional model followed by a biLSTM encoder.
2The first and last chunks would only have two consecutive
(sequence). We define the following “soft" cross-entropy loss:
L(π|S) =−1
l X
si∈S
X
πi∈{E,A,R}
y(πi) logp(πi),
(3) where, for a given sentence si ∈ S, y(πi) denotes its soft-label for decision.
We next explain how each soft-label y(πi) is estimated. To this end, we utilize a given summary quality metricr(S0)which can be used to evaluate the quality of any given summaryS0 (e.g., ROUGE (Lin,2004)). Overall, for a given text inputS withlsentences, there are3lpossible summariesS0 to consider. Letπ∗ = (π∗1, . . . , πl∗) denote the best decision sequence which results in the summary which maximizes r(·). For i ∈ {1,2, . . . , l}, let r(π¯ ∗1, . . . , πi−∗ 1, πi) denote the averager(·)value obtained by decision sequences that start with the prefix(π1∗, . . . , π∗i−1, πi). Based onπ∗, the soft labely(πi)is then calculated3 as follows:
y(πi) =
¯
r(π1∗, . . . , π∗i−1, πi) P
πj∈{E,A,R}r(π¯
∗
1, . . . , πi−∗ 1, πj) (4)
3 Evaluation
3.1 Dataset and Setup
We trained, validated and tested our approach using the non-annonymized version of the CNN/DailyMail dataset (Hermann et al., 2015). Following (Nallapati et al.,2016), we used the story highlights associated with each article as its ground truth summary. We further used the F-measure versions of 1, 2 and ROUGE-L as our evaluation metrics (Lin,2004).
The extractor and abstractor were trained similarly to (Chen and Bansal,2018) (including the same hyperparameters). The Editorial Network (hereinafter denotedEditNet) was trained according to Section 2.4, using the ADAM optimizer with a learning rate of 10−4 and a batch size of 32. Following (Dong et al., 2018;
Wu and Hu,2018a), we set the reward metric to be r(·) = αR-1(·) +βR-2(·) + γR-L(·); with α = 0.4,β = 1andγ = 0.5, which were further suggested by (Wu and Hu,2018a).
We further applied the Teacher-Forcing approach (Lamb et al., 2016) during training, where we considered the true-label instead of the
3Fori= 1we have:¯r(π∗ 1, . . . , π
∗
0, π1) = ¯r(π1).
Table 1: Quality evaluation using ROUGE
F-measure (ROUGE-1, ROUGE-2, ROUGE-L) on
CNN/DailyMail non-annonymized dataset
R-1 R-2 R-L
Extractive
Lead-3 40.00 17.50 36.20
SummaRuNNer(Nallapati et al.,2017) 39.60 16.20 35.30
EditNetE 38.43 18.07 35.37
Refresh(Narayan et al.,2018) 40.00 18.20 36.60
Rnes w/o coherence(Wu and Hu,2018b) 41.25 18.87 37.75
BanditSum(Dong et al.,2018) 41.50 18.70 37.60
Latent(Zhang et al.,2018) 41.05 18.77 37.54
rnn-ext+RL(Chen and Bansal,2018) 41.47 18.72 37.76
NeuSum(Zhou et al.,2018) 41.59 19.01 37.98
BERTSUM(Liu,2019) 43.25 20.24 39.63
Abstractive
Pointer-Generator(See et al.,2017) 39.53 17.28 36.38
KIGN+Prediction-guide(Li et al.,2018) 38.95 17.12 35.68
Multi-Task(EG+QG)(Guo et al.,2018) 39.81 17.64 36.54
EditNetA 40.00 17.73 37.53
rnn-ext+abs+RL(Chen and Bansal,2018) 40.04 17.61 37.59
RL+pg+cbdec(Jiang and Bansal,2018) 40.66 17.87 37.06
Saliency+Entail.(Pasunuru and Bansal,2018) 40.43 18.00 37.10
Inconsistency loss(Hsu et al.,2018) 40.68 17.97 37.13
Bottom-up(Gehrmann et al.,2018) 41.22 18.68 38.34
DCA(Celikyilmaz et al.,2018) 41.69 19.47 37.92
Mixed Extractive-Abstractive
EditNet 41.42 19.03 38.36
editor’s decision (including when updating gi at each stepiaccording to Eq.2). Following (Chen and Bansal,2018), we setm= 512andn= 512. We trained for20epochs, which has taken about 72 hours on a single GPU. We chose the best model over the validation set for testing. Finally, all components were implemented in Python3.6 using the pytorch0.4.1package.
3.2 Results
Table 1 compares the quality of EditNet with that of several state-of-the-art extractive-only or abstractive-only baselines. This includes the extractor (rnn-ext-RL) and abstractor ( rnn-ext-abs-RL) components of (Chen and Bansal,2018) that we utilized for implementingEditNet4.
We further report the quality of EditNetwhen it was being enforced to take an extract-only or abstract-only decision, denoted hereinafter as EditNetE and EditNetA, respectively. The comparison of EditNet to both EditNetE and
EditNetAvariants provides a strong empirical proof
that, by utilizing an hybrid decision approach, a
4The
better summarization quality is obtained.
Overall,EditNetprovides a highly competitive summary quality, where it outperforms most baselines. Interestingly,EditNet’s summarization quality is quite similar to that ofNeuSum(Zhou et al., 2018). Yet, while NeuSum applies an extraction-only approach, summaries generated by EditNet include a mixture of sentences that have been either extracted or abstracted.
Two models outperform EditNet, BERTSUM (Liu, 2019) and DCA (Celikyilmaz et al.,
2018). TheBERTSUMmodel gains an impressive accuracy, yet it is an extractive model that utilizes many attention layers running in parallel with millions of parameters (Devlin et al., 2019). DCA gains a comparable quality to EditNet, it outperforms on R-2 and slightly on R-1. The contextual encoder ofDCAis comprised of several LSTMlayers one on top of the other with varied number of agents (hyper-tuned) that transmit messages to each other. Considering the complexity of these models, and the slow down that can incur during training and inference, we think that EditNet still provides a useful, high quality and relatively simple extension on top of standard encoder aligned decoder architectures.
On average,56%and18%ofEditNet’s decisions were to abstract (A) or reject (R), respectively. Moreover, on average, per summary,EditNetkeeps only 33% of the original (extracted) sentences, while the rest (67%) are abstracted ones. This demonstrates that, EditNet has a high capability of utilizing abstraction, while being also able to maintain or reject the original extracted text whenever it is estimated to provide the best benefit for the summary’s quality.
4 Conclusions and Future Work
We have proposed EditNet – a novel alternative summarization approach that instead of solely applying extraction or abstraction, mixes both together. Moreover, EditNet implements a novel sentence rejection decision, allowing to “correct” initial sentence selection decisions which are predicted to negatively effect summarization quality. As future work, we plan to evaluate other alternative extractor-abstractor configurations and try to train the network end-to-end. We further plan to explore reinforcement learning (RL) as an alternative decision making approach.
References
Asli Celikyilmaz, Antoine Bosselut, Xiaodong He, and Yejin Choi. 2018. Deep communicating agents for
abstractive summarization. In Proceedings of the
2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long
Papers), pages 1662–1675.
Yen-Chun Chen and Mohit Bansal. 2018. Fast
abstractive summarization with reinforce-selected sentence rewriting. In Proceedings of the 56th Annual Meeting of the Association for
Computational Linguistics (Volume 1: Long Papers),
pages 675–686. Association for Computational Linguistics.
Sumit Chopra, Michael Auli, and Alexander M. Rush.
2016. Abstractive sentence summarization with
attentive recurrent neural networks. InProceedings of the 2016 Conference of the North American Chapter of the Association for Computational
Linguistics: Human Language Technologies, pages
93–98. Association for Computational Linguistics.
Arman Cohan, Franck Dernoncourt, Doo Soon Kim, Trung Bui, Seokhwan Kim, Walter Chang, and
Nazli Goharian. 2018. A discourse-aware attention
model for abstractive summarization of long documents. InProceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language
Technologies, Volume 2 (Short Papers), pages 615–
621. Association for Computational Linguistics.
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and
Kristina Toutanova. 2019. Bert: Pre-training
of deep bidirectional transformers for language
understanding. In Proceedings of the 2019
Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short
Papers), pages 4171–4186.
Yue Dong, Yikang Shen, Eric Crawford, Herke van
Hoof, and Jackie Chi Kit Cheung. 2018.Banditsum:
Extractive summarization as a contextual bandit. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Brussels,
Belgium, October 31 - November 4, 2018, pages
3739–3748.
Sebastian Gehrmann, Yuntian Deng, and Alexander
Rush. 2018. Bottom-up abstractive summarization.
InProceedings of the 2018 Conference on Empirical
Methods in Natural Language Processing, pages
4098–4109. Association for Computational
Linguistics.
Han Guo, Ramakanth Pasunuru, and Mohit Bansal.
2018. Soft layer-specific multi-task summarization
1: Long Papers), pages 687–697. Association for Computational Linguistics.
Karl Moritz Hermann, Tomáš Koˇciský, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa
Suleyman, and Phil Blunsom. 2015. Teaching
machines to read and comprehend. InProceedings of the 28th International Conference on Neural
Information Processing Systems - Volume 1,
NIPS’15, pages 1693–1701, Cambridge, MA, USA. MIT Press.
Wan-Ting Hsu, Chieh-Kai Lin, Ming-Ying Lee, Kerui
Min, Jing Tang, and Min Sun. 2018. A unified
model for extractive and abstractive summarization using inconsistency loss. In Proceedings of the 56th Annual Meeting of the Association for
Computational Linguistics (Volume 1: Long Papers),
pages 132–141. Association for Computational Linguistics.
Yichen Jiang and Mohit Bansal. 2018.
Closed-book training to improve summarization encoder memory. InProceedings of the 2018 Conference on
Empirical Methods in Natural Language Processing,
pages 4067–4077. Association for Computational Linguistics.
Alex M Lamb, Anirudh Goyal ALIAS PARTH GOYAL, Ying Zhang, Saizheng Zhang, Aaron C
Courville, and Yoshua Bengio. 2016. Professor
forcing: A new algorithm for training recurrent
networks. In Advances In Neural Information
Processing Systems, pages 4601–4609.
Chenliang Li, Weiran Xu, Si Li, and Sheng Gao.
2018. Guiding generation for abstractive text
summarization based on key information guide network. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language
Technologies, Volume 2 (Short Papers), pages 55–60.
Association for Computational Linguistics.
Chin-Yew Lin. 2004. Rouge: A package for automatic
evaluation of summaries. In Text summarization
branches out: Proceedings of the ACL-04 workshop,
volume 8. Barcelona, Spain.
Peter J Liu, Mohammad Saleh, Etienne Pot, Ben Goodrich, Ryan Sepassi, Lukasz Kaiser, and
Noam Shazeer. 2018. Generating wikipedia by
summarizing long sequences. arXiv preprint
arXiv:1801.10198.
Yang Liu. 2019. Fine-tune bert for extractive
summarization.arXiv preprint arXiv:1903.10318.
Ramesh Nallapati, Feifei Zhai, and Bowen Zhou. 2017.
Summarunner: A recurrent neural network based sequence model for extractive summarization of documents. InProceedings of the Thirty-First AAAI Conference on Artificial Intelligence, February 4-9,
2017, San Francisco, California, USA., pages 3075–
3081.
Ramesh Nallapati, Bowen Zhou, Cícero Nogueira dos Santos, Çaglar Gülçehre, and Bing Xiang. 2016.
Abstractive text summarization using sequence-to-sequence rnns and beyond. InProceedings of the 20th SIGNLL Conference on Computational Natural Language Learning, CoNLL 2016, Berlin, Germany,
August 11-12, 2016, pages 280–290.
Shashi Narayan, Shay B. Cohen, and Mirella
Lapata. 2018. Ranking sentences for extractive
summarization with reinforcement learning. In
Proceedings of the 2018 Conference of the
North American Chapter of the Association for
Computational Linguistics: Human Language
Technologies, NAACL-HLT 2018, New Orleans, Louisiana, USA, June 1-6, 2018, Volume 1 (Long
Papers), pages 1747–1759.
Ramakanth Pasunuru and Mohit Bansal. 2018.
Multi-reward reinforced summarization with saliency and entailment. InProceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language
Technologies, Volume 2 (Short Papers), pages 646–
653. Association for Computational Linguistics.
Romain Paulus, Caiming Xiong, and Richard Socher.
2017. A deep reinforced model for abstractive
summarization.CoRR, abs/1705.04304.
Abigail See, Peter J. Liu, and Christopher D. Manning.
2017. Get to the point: Summarization with
pointer-generator networks. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, ACL 2017, Vancouver,
Canada, July 30 - August 4, Volume 1: Long Papers,
pages 1073–1083.
Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. 2014.
Sequence to sequence learning with neural networks.
InProceedings of the 27th International Conference
on Neural Information Processing Systems - Volume
2, NIPS’14, pages 3104–3112, Cambridge, MA,
USA. MIT Press.
Jiwei Tan, Xiaojun Wan, and Jianguo Xiao. 2017.
Abstractive document summarization with a graph-based attentional neural model. In Proceedings of the 55th Annual Meeting of the Association for
Computational Linguistics (Volume 1: Long Papers),
pages 1171–1181. Association for Computational Linguistics.
Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly.
2015. Pointer networks. In Advances in Neural
Information Processing Systems, pages 2692–2700.
Yuxiang Wu and Baotian Hu. 2018a. Learning to
extract coherent summary via deep reinforcement learning. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence, (AAAI-18), the 30th innovative Applications of Artificial Intelligence (IAAI-18), and the 8th AAAI Symposium on Educational Advances in Artificial Intelligence (EAAI-18), New Orleans, Louisiana, USA, February
Yuxiang Wu and Baotian Hu. 2018b. Learning to extract coherent summary via deep reinforcement learning. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence, (AAAI-18), the 30th innovative Applications of Artificial Intelligence (IAAI-18), and the 8th AAAI Symposium on Educational Advances in Artificial Intelligence (EAAI-18), New Orleans, Louisiana, USA, February
2-7, 2018, pages 5602–5609.
Xingxing Zhang, Mirella Lapata, Furu Wei, and
Ming Zhou. 2018. Neural latent extractive
document summarization. In Proceedings of the 2018 Conference on Empirical Methods in Natural
Language Processing, pages 779–784. Association
for Computational Linguistics.
Qingyu Zhou, Nan Yang, Furu Wei, Shaohan Huang,
Ming Zhou, and Tiejun Zhao. 2018. Neural
document summarization by jointly learning to score and select sentences. In Proceedings of the 56th Annual Meeting of the Association for
Computational Linguistics (Volume 1: Long Papers),
